مشاركة عبر

Azure AI Video Indexer enabled by Arc (preview)

  • مقالة

Azure AI Video Indexer enabled by Arc is an Azure Arc extension enabled service that runs video and audio analysis, and generative AI on edge devices. The solution is designed to run on Azure Arc enabled Kubernetes and supports many video formats, including MP4 and other common formats. It supports several languages in all basic audio-related models. It assumes that one Video Indexer resource is mapped to one extension.

If you aren't already familiar with Azure AI Video Indexer, it's recommended that you familiarize yourself with the cloud service first.

Additionally, before you start working with Azure AI Video Indexer enabled by Arc, review the transparency note to understand usage restrictions.

Important

To successfully deploy the Azure AI Video Indexer extension, it is mandatory that your Azure subscription id is approved in advance. You must first sign up using this form.

What is Azure Arc and Azure Arc-enabled Kubernetes?

Azure Arc simplifies governance and management of complex environments that extend across data centers, multiple clouds, and edge by delivering a consistent multi-cloud and on-premises management platform.

Azure Arc-enabled Kubernetes allows you to attach Kubernetes clusters running anywhere so that you can manage and configure them in Azure. By managing all of your Kubernetes resources in a single control plane, you can enable a more consistent development and operation experience to run cloud-native apps anywhere and on any Kubernetes platform.

When the Azure Arc agents are deployed to the cluster, an outbound connection to Azure is initiated, using industry-standard SSL to secure data in transit.

Once clusters are connected to Azure, they're represented as their own resources in Azure Resource Manager (ARM), and they can be organized using resource groups and tagging.

See these articles to understand more about Azure Arc and Azure Arc-enabled Kubernetes.

What is an Azure Arc extension?

Virtual machine (VM) extensions are small applications that provide post-deployment configuration and automation tasks on Azure VMs. For example, if a virtual machine requires software installation, anti-virus protection, or to run a script in it, a VM extension can be used. To understand more about extensions, see Virtual machine extension management with Azure Arc-enabled servers.

The Azure AI Video Indexer extension installs and deploys Azure AI Video indexer to the Kubernetes cluster.

All Azure AI Video Indexer enabled by Arc only supports Azure Resource Manager (ARM) accounts. ARM operations are decoupled from video insight operations. This design allows you to perform analysis on your edge devices without the need to upload your media assets to Azure.

Azure AI Video Indexer enabled by Arc doesn't support classic accounts. For more information about the retirement of classic accounts, see Preparing for AMS retirement: VI migration and updating guide

The extension is supported in direct connection mode scenarios only. Control plane information is sent to the cloud, for example, monitoring, usage. New extension versions are downloaded from the cloud. No customer data, such as what videos were indexed, is sent from the edge location to the cloud.

Language models

The Phi 3 language model is included and automatically connected with your VI extension. You can start using it immediately. For more information about using language models with VI see:

See also the transparancy note for textual summarization with Vi enabled by Arc for hardware requirements, limitations, and known issues.

Use cases

  • Data governance – You can bring the AI to the content instead of vice versa. Use Azure AI Video Indexer enabled by Arc when you can’t move indexed content from on-premises to the cloud due to:
    • regulation.
    • architecture decisions.
    • data store being too large, making lift and shift a significant effort.
  • On-premises workflow – Your indexing process is part of an on-premises workflow, and you want to lower the indexing duration latency affecting the flow.
  • Pre-indexing – You want to index before uploading the content to the cloud. To create clarity, you can presort your on-premises video and/or audio archive, and then only upload it for standard and/or advanced indexing in the cloud.

Example deployment

The following block diagram shows the Azure AI Video Indexer extension running on Azure Arc. There are three types:

  1. Store type A uses both vision and audio presets.
  2. Store type B uses only vision presets. It also has a custom model. For more information about using a custom model with Azure AI Video Indexer enabled by Arc, see Bring Your Own AI model.
  3. Store C uses only audio presets.

The extension is stored on each edge device and each device is associated with a single Azure AI Video Indexer account that interfaces with Azure Arc and the cloud.

VI Arc block diagram

Supported AI presets

Azure AI Video Indexer enabled by Arc supports the following indexing presets:

Model Basic Video Basic Audio Basic Video & Audio
Transcription ✔️ ✔️
Translation ✔️ ✔️
Captioning ✔️ ✔️
Key frame detection ✔️ ✔️
OCR ✔️ ✔️
Object detection ✔️ ✔️
Scene detection ✔️ ✔️
Shot detection ✔️ ✔️
Summarization ✔️ ✔️

Minimum hardware requirements

Video Indexer enabled by Arc is designed to run on any Arc enabled Kubernetes environment.

The following list is the minimum and recommended requirements if the extension contains single language support. If you install multiple speech and translation containers with several languages, increase the hardware requirements accordingly.

Note

These are minimum requirements for a production environment. At least a 2-node cluster is recommended for high availability and scalability. The recommended settings refer to cluster wide settings, so for example, if you have 2 nodes, each node should have 16 cores and 32 GB of RAM. We recommend creating a dedicated node-pool / auto-scaling groups to host the VI Solution.

Configuration VM Count Node CPU Cores Count Node Ram Node Storage Remarks
Minimum 1 32 Cores 64 GB 50 GB Storage needs to support ReadWriteMany Storage Class
Recommended 2 48-64 Cores 256 GB 100 GB Storage needs to support ReadWriteMany Storage Class

Minimum software requirements

Component Minimum Requirements
Operating System Ubuntu 22.04 LTS or any Linux Compatible OS
Kubernetes 1.26
Azure CLI 2.48.0

Supported input formats and codecs

Video formats

  • AVI (.avi)
  • FLV (with H.264 and AAC codecs) (.flv)
  • ISMV (.isma, .ismv)
  • Matroska/WebM (.mkv)
  • MP4 (.mp4, .m4a, .m4v)
  • MXF (.mxf)
  • MPEG2-TS
  • QuickTime (.mov)
  • WAVE/WAV (.wav)
  • Webm
  • Windows Media Video (WMV)/ASF (.wmv, .asf)

Video codecs

Here is your alphabetized list:

  • AVC 8-bit/10-bit, up to 4:2:2, including AVCIntra
  • Digital video (DV) (in AVI files)
  • DVCPro/DVCProHD (in MXF container)
  • HEVC/H.265
  • MPEG-1
  • MPEG-2 (up to 422 Profile and High Level; including variants such as Sony XDCAM, Sony XDCAM HD, Sony XDCAM IMX, CableLabs®, and D10)
  • MPEG-4 Part 2
  • VC-1/WMV9

Audio codecs up to two tracks

  • AAC (AAC-LC, AAC-HE, and AAC-HEv2)
  • FLAC
  • MPEG Layer 2
  • MP3 (MPEG-1 Audio Layer 3)
  • VORBIS
  • WAV/PCM
  • Windows Media Audio

Supported languages

  • Arabic (Saudi Arabia)
  • Arabic Egypt
  • Chinese (Simplified)
  • English (US)
  • French
  • German
  • Italian
  • Spanish

Bring your own model

Azure AI Video Indexer enabled by Arc also supports bringing your own model. See the Bring Your Own Model (BYO) article for details.

Limitations

  • The supported file size for indexing is up to 2 GB.
  • Upgrading the extension:
    • Extension support applies for the latest version only.
    • We recommend setting that auto-upgrade property to true. This setting keeps the extension up to date.
    • If the auto upgrade setting is set to false, the version upgrade should be done incrementally. Jumping between versions can cause indexing processes to fail.
  • After the extension installation or upgrade, expect the first index\translation process duration to be longer. The longer duration is due to AI model image download. The duration varies depending on network speed.
  • Only one Video Indexer extension can be deployed per Arc enabled Kubernetes cluster.
  • The cluster's volume performance (based on storage class) has significant influence on the turnover duration of the indexing job especially since the frame extraction is writing all frames into the volume).
  • You can use only cloud account access tokens obtained via the Azure portal. Cloud video access tokens aren't supported, but with the API, extension access tokens are available and we support all types.
  • Video error messages aren't stored due to memory limitations.